OCPI·LIVE
H100from $1.95H100from $1.66H100from $2.59H200from $3.00H200from $3.79A100from $1.00A100from $1.99A100from $1.35A100from $1.35B200from $3.74B300from $6.94L40Sfrom $0.69RTX 4090from $0.34RTX 5090from $0.65H100from $1.95H100from $1.66H100from $2.59H200from $3.00H200from $3.79A100from $1.00A100from $1.99A100from $1.35A100from $1.35B200from $3.74B300from $6.94L40Sfrom $0.69RTX 4090from $0.34RTX 5090from $0.65
Index / Compare
Cloud GPU Price Comparison

B300 vs H200

B300 SXM 288GB vs H200 SXM 141GB — current cloud rental cost per hour, normalized to USD per GPU across providers. H200 SXM 141GB is currently the cheaper option — about 46% less on the headline comparison metric.

B300 SXM 288GB
NVIDIA Blackwell Ultra · SXM
$7.40/GPU/hr
single index provider · not a median
range $6.94–$16.82 · 4 providers
cheapest: RunPod $6.94
↗ Full B300 SXM 288GB prices
H200 SXM 141GB
NVIDIA Hopper · SXM
$3.99/GPU/hr
single index provider · not a median
range $3.00–$7.46 · 3 providers
cheapest: Shadeform $3.00
↗ Full H200 SXM 141GB prices

B300 vs H200: memory and workload context

spec context · not a performance benchmark
ContextB300 SXM 288GBH200 SXM 141GB
ArchitectureNVIDIA Blackwell UltraNVIDIA Hopper
Memory288GB HBM3e141GB HBM3e
Common workloadsfrontier model training, reasoning inference, memory-intensive AI workloadslarge-context inference, memory-bound training, serving larger open-weight models
Price interpretationB300 is a distinct 288GB Blackwell Ultra GPU. It is not combined with GB300 systems, and early cloud coverage should be read with the provider count.H200 adds substantially more memory than H100, so price comparisons should consider whether the workload is memory-bound.

B300 vs H200 price table

USD per GPU per hour · live provider data
MetricB300 SXM 288GBH200 SXM 141GB
Index median$7.40/hr$3.99/hr
Community marketplace median$6.94/hr$3.59/hr
Cheapest current row$6.94 · RunPod$3.00 · Shadeform
Provider coverage4 providers · 18 listings3 providers · 21 listings
Best use of this pageUse this as a price comparison and sourcing snapshot. Performance, memory size, software support, and availability still need to be checked with the provider.
“

Listed price ≠ transaction price

These are published rates, normalized per GPU per hour and labeled by billing mode. The index median uses eligible reliable-tier providers only. Read the methodology →